On the observational implications of Knightian

uncertainty
 On the Observational Implications of Knightian
Uncertainty
Kevin A. Hassett and Weifeng Zhong
American Enterprise Institute

December 21, 2016

Abstract
We develop a model of a prediction market with ambiguity and derive
testable implications of the presence of Knightian uncertainty. Our model can
explain two commonly observed empirical regularities in betting markets: the
tendency for longshots to win less often than odds would indicate and the ten-
dency for favorites to win more often. Using historical data from Intrade, we
further present empirical evidence that is consistent with the predicted presence
of Knightian uncertainty. Our evidence also suggests that, even with informa-
tion acquisition, the Knightian uncertainty of the world may be not learnable
to the traders in prediction markets.

Keywords: ambiguity, Knightian uncertainty, prediction market, maxmin

preferences

1 Introduction
At least since the work of Knight (1921), economists have understood that economic
agents may behave dierently in risky circumstances, where outcomes are random
but governed by known probabilities, as opposed to uncertain circumstances, where
risks are unknown. Ellsberg (1961) provides examples that highlight the tendency
for some decision makers to be averse to the presence of Knightian uncertainty or,
ambiguity.
Address: American Enterprise Institute, 1789 Massachusetts Avenue, NW, Washington, DC
20036. Email addresses: khassett@aei.org and weifeng.zhong@aei.org. We thank Jon Hartley, Cody
Kallen, and Joe Sullivan for excellent research assistance.

1
 In recent years, there has been an explosion of theoretical work developing models
that incorporate ambiguity aversion, building o of the seminal contribution of
Gilboa and Schmeidler (1989). In the literature to date, Knightian uncertainty has
been a factor inserted in a model that could possibly explain puzzling observations.
It has served a role analogous to that of dark matter in cosmological models, lurking
behind the scenes to explain observed phenomena, never being directly observed.
At the same time, a rich literature has evolved exploring the e ciency of betting
and prediction markets that price specic events. Following on the early work of
Kahneman and Tversky (1979) and Asch, Malkiel and Quandt (1982), the ability of
these markets to predict future events has been studied extensively, and a number
of empirical anomalies have been identied.
In this paper, we extend the theoretical literature and connect it to the prediction-
market application. In so doing, we develop more direct observable implications of
the presence of Knightian uncertainty than has been achieved previously in the
literature, and a method to test for its presence.
While we below will formally derive a model that suggests our test, the intuition
of our approach is quite straightforward and can be illustrated using an example
from Ellsberg (1961). Suppose that we have two urns. In one urn, we have 50 black
balls and 50 red balls. In another urn the Knightian urn we have 100 balls,
but we have no information regarding the proportions. A subject is oered a game.
If she pulls a black ball out of the urn, she wins $1. If she pulls a red ball out she
wins nothing. The literature has documented a tendency for individuals to prefer
the urn with the known probabilities, suggesting that they exhibit the aversion to
ambiguity discussed above.
Suppose that an econometrician could observe games played with both of the
urns in Ellsbergs game. With a number of repeated trials, the sample propor-
tions from the rst urn would fairly rapidly indicate an estimate that the binomial
probability of victory is 50 percent. With enough data, one would say that with
great condence. On the other hand, if one observed repeated play with the second,
Knightian urn which, after all, has some number of black balls in it, then the sam-
ple proportion would also converge to an estimated binomial probability, but that
probability would not necessarily be 50 percent.
The observation that motivates this paper stems from this thought experiment.
Given a market derived ex ante probability of a binary event, as one frequently
observes in betting markets, there will naturally be circumstances where informa-

2
tion is extremely solid, and odds are quite far from 50 percent. There will also be
situations where information suggests there is an even match (as with a coin ip),
and the contract suggests there is close to a 50 percent chance of either outcome.
This often happens, for example, in presidential futures markets in the U.S. after the
conventions are over. But it is also possible that there are contracts that suggest
that the odds of either outcome are 50 percent because the event is shrouded in
ambiguity. If we were to estimate the ex post sample proportions from just these
contracts with ex ante 50 percent probabilities, then they could, as in the Ellsberg
example above, be anything. If we were to estimate the ex post sample proportions
of the high information contracts with probabilities far from 50 percent, the pro-
portions and ex ante probabilities should, if markets are e cient, align. But close
to 50 percent, they might not, and if they do not, it is indication of the presence of
Knightian uncertainty. Thus, the pattern by which the relationship between ex post
proportions and ex ante probabilities deviates from the 45 degree line becomes infor-
mative regarding the presence of Knightian uncertainty. We also discuss the extent
to which learning can occur in markets over time. If Knightian uncertainty induces
knowledge acquisition, then the relationship between proportions and probabilities
will evolve as a market matures, a possibility we explore in the paper.
The next subsection briey reviews the literature. In Sections 2 and 3, we draw
on the work of Gilboa and Schmeidler (1989) and Dow and Werlang (1992) and
develop a model that suggests that the pattern described by our intuitive exam-
ple would emerge in a market inuenced by the present of signicant Knightian
uncertainty. In Section 4, we provide some high-level evidence that the relation-
ship between ex post proportions and ex ante probabilities is consistent with the
predictions of our model. Section 5 concludes.

1.1 Literature Review

This paper draws from two dierent strands in the literature. First, theorists have
made remarkable strides in recent years incorporating Knightian uncertainty and
ambiguity aversion into models of nancial markets.
These models have, according to an exhaustive recent review, implications for
portfolio choice and asset pricing that are very dierent from those of SEU (subjec-
tive expected utility theory) and that help to explain otherwise puzzling features of

3
the data.1 Ambiguity aversion could help explain the tendency of markets to stop
operating during nancial crises, for prices to not be completely informative, and
even for there to be bank runs.2
This branch of the literature has focused on nancial markets in general. At the
same time, an equally impressive literature has emerged exploring the functioning
of prediction markets, which, for the most part, price in the probability of specic
binary events. As Thaler and Ziemba (1988) rst noted, these prediction markets
may be a better laboratory to test cutting edge theories, as they contain contracts
with known durations, and observable discrete events that stop the trading. While
an equity might live on virtually forever, a presidential election future has a specic
end date, and its ability to forecast the outcome can be precisely evaluated.
This second literature has advanced both empirically and theoretically. On the
theoretical side, Manski (2004) rst illustrated that the beliefs of bettors may not
necessarily yield a market-based probability. More recently, Wolfers and Zitzewitz
(2006) identify the conditions under which prediction-market prices coincide with
bettorsmean beliefs about probabilities. On the empirical side, prediction markets
have been found to be informative regarding the odds of events occurring. Berg
et al. (2008), for example, nd that the Iowa Electronic Markets outperformed
polls in predicting election outcomes. At the same time, markets have been found
to exhibit a favorite-longshot bias, with favorites outperforming their odds, and
longshots underperforming (see, e.g., Cain, Law and Peel, 2000). A number of
possible explanations for this pattern include insider trading (Shin, 1992), risk loving
behavior (Weitzman, 1965), and imperfect ability to process information (Snowberg
and Wolfers, 2010).
The connection of these two literatures seems quite promising, as betting markets
often exist for events, such as Brexit or elections, for which Knightian uncertainty
may well be present. Since they also have nite and determinate life spans, they
also allow the econometrician the ability to evaluate their performance ex post. We
now turn to illustrating the utility of this approach.
1
See Epstein and Schneider (2010), p. 315.
2
See Caballero and Krishnamurthy (2008), Caballero and Simsek (2013), Guidolin and Rinaldi
(2010), Routledge and Zin (2009), and Uhlig (2013).

4
2 A Model of Prediction Market with Ambiguity
2.1 Setup
Events and Contracts. Consider a prediction market for the occurrence of a bi-
nary event. There are two all-or-nothing contracts corresponding to the two possible
realizations. One contract pays $1 if event A occurs and $0 otherwise, while the
other contract pays $1 if the complementary event Ac occurs and $0 otherwise. Let
denote the price of contract A. No-arbitrage condition dictates that, in equilibrium,
the price of contract Ac be 1 .

Traders. There is a continuum I of competitive traders, each endowed with ho-

mogeneous initial wealth w. The net position on contract A held by trader i is
denoted by xi 2 R.3 Given price , the nal wealth wi of trader i is
(
w + (1 ) xi if event A occurs,
wi =
w xi if event Ac occurs.

All traders have log utility of their nal wealth: u (wi ) = ln wi .

Beliefs and Ambiguity. Suppose trader i has a subjective belief that event A
occurs with probability qe 2 [0; 1]. Then, the subjective expected utility of trader i
from holding position xi at price is given by

U ( ; xi ; qe) = qe ln (w + (1 ) xi ) + (1 qe) ln (w xi ) :

However, ambiguity exists, for traders may be uncertain about how likely event
A is to occur. We follow Gilboa and Schmeidler (1989) and model ambiguity using
the multiple-prior framework. Specically, suppose each trader i considers every
probability qe 2 [qi ; qi + ], where 0, an admissible probability that governs
the realization of the binary event. Under this framework, qi represents the mean
belief of trader i, while is interpreted as a measure of ambiguity. Given price
3
In practice, trader i can long and/or short contract A and/or contract Ac ; but some strategies
are mathematically equivalent. For example, holding mi > 0 units of contract A and ni > mi > 0
units of contract Ac would be equivalent to holding mi units of cash, 0 unit of contract A , and
ni mi > 0 units of contract Ac : Therefore, without loss of generality, we let a single decision
variable xi = mi ni (which could be positive, zero, or negative) represent the net position held
by trader i:

5
 , trader i chooses position xi to maximize the minimum that is, the worst-case
scenario of all her admissible, subjective expected utilities:

max min U ( ; xi ; qe) : (1)

xi 2R qe2[qi ;qi + ]

Traders are heterogeneous in mean belief. Let the distribution of tradersmean

beliefs be characterized by a cumulative distribution function F over interval [ ; 1 ].
That is, for the most pessimistic trader, the worst-case belief that A occurs is prob-
ability 0 while, for the most optimistic trader, the best-case belief that A occurs is
probability 1.

2.2 Optimal Demand and Portfolio Inertia

Solving the inner minimization reduces the optimization problem (1) to

max U ( ; xi ; qi sgn (xi ) ) ;

xi 2R

where sgn ( ) is an indicator function that takes the sign of its argument.
The intuition behind the above expression is straightforward. If trader i has a
positive position on contract A, then the worst-case scenario would be that event A
occurs with probability qi , the lower bound. Similarly, if the position of trader i
is negative, then, in the worst-case scenario, event A occurs with the upper-bound
probability, qi + .
Solving the maximization problem gives the optimal (net) demand for contract
A by trader i,
8 qi
>
>
< (1 )w if 2 [0; qi );
x ( ; qi ) = 0 if 2 [qi ; qi + ] ; (2)
>
>
: qi +
(1 )w if 2 (qi + ; 1] ;

as a function of price and mean belief. Therefore, trader i longs contract A when
the price is lower than her most pessimistic belief, and shorts contract A when the
price is higher than her most optimistic belief. For any price in the intermediate
range [qi ; qi + ], trader i does not participate in the prediction market the
phenomenon of portfolio inertia.
That portfolio inertia arises when investors have maxmin preferences is well
known in the nance literature since the work by Dow and Werlang (1992). The

6
setup of this model replicates this phenomenon in the context of prediction markets.
In particular, for each trader, the size of price region at which portfolio inertia occurs
is given by 2 . In other words, the higher the degree of ambiguity, the more inertial
the tradersportfolios.

2.3 Equilibrium
Given price for contract A and distribution function F of tradersmean beliefs,
the aggregate (net) demand for the contract is given by
Z 1
XF ( ) = x ( ; q) dF (q) : (3)

The prediction market is in equilibrium when the aggregate demand for con-
tract A equals zero, that is, XF ( ) = 0. The following proposition establishes the
equilibrium price.4

Proposition 1 Given distribution function F , the equilibrium price F is such that

Z
F+
F = EF (q) + F (q) dq :
F

When ambiguity is absent (i.e., = 0) the prediction market aggregates the

wisdom of crowds:
F j =0 = EF (q) :

That is, the equilibrium price of contract A corresponds to the average of traders
mean beliefs about the occurrence of event A.
In the presence of ambiguity, however, the prediction market does not necessarily
aggregate the wisdom of crowds. In particular, it aggregates the wisdom of crowds
R +
if and only if the distribution function F is such that F F (q) dq = . The next
F
proposition shows that the situations in which such equality happens to hold are
topologically rare.

Proposition 2 The prediction market rarely aggregates the wisdom of crowds.

Formally, let be the space of probability distributions over [ ; 1 ], endowed with
the weak topology. Then, the subset of probability distributions such that the equilib-
rium price equals the average of traders mean beliefs is nowhere dense in .
4
We relegate all proofs to Appendix A.

7
 Propositions 1 and 2 together suggests that the presence of ambiguity renders
the prediction market ineective in aggregating the beliefs held by heterogeneous
traders.
The reason the prediction market fails to aggregate beliefs is that ambiguity
deters market participation by traders and, hence, the beliefs held by them.

Proposition 3 The equilibrium quantity of trades is strictly decreasing in the degree

of ambiguity.

Proposition 3 is a direct consequence of portfolio inertia. As the degree of ambi-

guity increases, the inaction rangeof each trader i, [qi ; qi + ], becomes wider.
Since each trader is more likely to stay put in a more ambiguous environment, the
aggregate trades must be fewer as well. This result is reminiscent of well-known
models of ambiguity in nancial economics (e.g., Caballero and Krishnamurthy,
2008; Guidolin and Rinaldi, 2010; and Routledge and Zin,2009), which suggest that
a signicant increase in Knightian uncertainty may contribute to liquidity hoarding
and market breakdown.

3 Testable Implications
The previous section has derived the equilibrium results under ambiguity. However,
since the degree of ambiguity is not observable, those results cannot be tested di-
rectly. In this section, we impose more information structures on the model and
derive implications that are testable with prediction-market data.

3.1 Information Structures

Suppose the true probability that event A occurs is given by p 2 [ ; 1 ]. No trader
knows about p for certain. However, there is a mass m 2 (0; 1) of traders whose
mean beliefs exactly equal p, while all the other uninformed traders mean beliefs
are continuously distributed over [ ; 1 ]. Assumption 1 embeds these additional
structures into the distribution function F .

Assumption 1 The distribution function F takes the following form:

(
(1 m) F (q) if q 2 [ ; p) ;
F (q)
(1 m) F (q) + m if q 2 [p; 1 ];

8
where F is some continuous distribution function of q over [ ; 1 ].

The functional form of F is left unspecied. We let denote the integral of F ,

Rq
i.e., (q) F (q 0 ) dq 0 .
One could provide a micro-foundation for this setup by assuming that the mass m
of traders have the correct mean beliefs because they have received private signals
informative of the true probability, while all other traders have received no such
signals. With such a micro-foundation, the traders (multiple) beliefs should be
interpreted as their (multiple) posteriors. In this section, we adopt a reduced-form
approach and build these details directly into the distribution function F .
The essence of this setup is that a non-negligible fraction of the traders hold
beliefs that are informative of the true state of the world. Therefore, since these
partially informed traders may or may not participate in the market, their beliefs
may or may not be reected in the equilibrium price.

3.2 Implications
Since the distribution function F is given and parameterized by the true probability
p, applying Proposition 1 allows us to solve for the equilibrium price as a function
of p, as shown below.

Proposition 4 Under Assumption 1, the equilibrium price (p) is:

1. continuous, with ( )> and (1 )<1 ;

2. such that (p) = b for any p 2 [b ; b + ];

3. strictly increasing for p 2

= [b ; b + ];

where b is identied by b (b + ) + (b )=1 2 (1 ).

Figure 1 plots the equilibrium price in a p diagram, where the true probability
1
p=( ) ( ) is a correspondence of the equilibrium price . Specically, it attains
a non-singleton set value when = b, with the size of that set equal to 2 .
The most important feature of the equilibrium (part 2) is that there exists a
range of true probabilities, [b ; b + ], within which the market price is not at all
responsive to any change in the underlying state of the world. That is, 0 (p) =0
for any p in that range. Instead of prediction, the prediction market simply assigns

9
 Figure 1: Prediction Market Equilibrium in the p Diagram.

an uninformative number b the mid-point of the range [b ; b + ] as the price.

The reason for this result is straightforward: Since the traders who hold private
information about p are not trading, what exactly those traders know about the
true state of the world must not be reected in the market price.
Outside the range [b ; b + ], however, the prediction market works (part 3).
Specically, if all parameters of the model were known, one would be able to infer the
true probability p from the equilibrium market price (p). Moreover, the higher
the true probability, the higher the price.
Part 1 of the proposition also shows that, for a true probability that is very
high (near 1 ) or very low (near ), the equilibrium price exhibits a favorite-
longshot bias commonly observed in the literature (e.g., Cain, Law, and Peel, 2000):
favorite events are under-priced while longshot events are over-priced. The intuition
is as follows. For a longshot event where p = , for example, if the market price
was as low as , that would imply all traders mean beliefs were greater than the
prevailing price and, hence, all traders would long the contract, which cannot be
an equilibrium. Therefore, the equilibrium price of a longshot must be signicantly
larger than the longshots odds.

10
 Note that, although the degree of ambiguity, , is not directly observable in real-
ity, Proposition 4 yields implications of the presence of ambiguity that are testable
with prediction-market data. Suppose an econometrician could conduct a large num-
ber of repeated trials for each value of the true probability. Then, with enough data,
the ex post sample proportion, denoted by P , would converge to the corresponding
true probability, p. It follows that the estimated relationship between P and the
1
market price, , would converge to the graph of the correspondence p = ( ) ( ).
As in Figure 1, such an ideal test would show a big jump at price level = b,
where the magnitude of the jump is an empirical measure of the degree of ambigu-
ity. Moreover, the graph should show an explicit pattern, with P being generally
below the 45-degree line below but close to b, and above it just thereafter.
Before conducting the test, the econometrician may not know where the jump
would appear, because b, given by

b (b + ) + (b )=1 2 (1 );

depends on the distribution of mean beliefs among all prediction-market traders. It

follows from the above equation that b would be smaller than 0:5 if F is skewed
towards the lowest mean belief , and larger than 0:5 if F is skewed towards the
highest mean belief 1 . But when F is symmetrically distributed over [ ; 1 ],
b would be equal to 0:5, which is the following corollary.

Corollary 5 Under Assumption 1,

b = 0:5

if F is a symmetric distribution function over [ ; 1 ] (i.e., F (1 x) = 1 F (x)

for any x 2 [ ; 1 ]).

In practice, the empirical chart would precisely follow Figure 1 with the jump at
0:5 in the case of symmetry, but not if asymmetries were present. But even if one
might expect skewness to be present for some contracts but not others, the range for
the crossover point could be scattered about the neighborhood of 0:5. Aggregation
of a large number of contracts, therefore, could push the average b to be in the
neighborhood of 0:5. But since each contract would exhibit a similar (if slightly
shifted) pattern, the overall pattern should loosely follow Figure 1 if Knightian un-
certainty is important in these markets, even though some reect symmetry whereas

11
others do not.
Accordingly, the theory suggests that a regression of P on would, because of
market e ciency when the market is fully functioning, reproduce a 45-degree line,
except for in the neighborhood of 0:5 (or an adjacent crossover point), where one
would expect to see signicant outliers below the line to the left of the crossover
point, and signicant positive outliers to the right. The pattern of outliers around
the regression line would be an indication that Knightian uncertainty is a factor in
the market, and would be consistent with the intuition provided in the introduction.

4 Empirical Evidence
In this section, we provide some high-level evidence that is consistent with the
theoretical predictions.5
We use the historical data from Intrade, a popular online prediction platform
which operated from 2003 to 2013. The platform hosted prediction contracts across
wide-ranging categories of events, such as business (e.g., whether the CEO of a
certain company would step down), current events (e.g., which city would host the
Olympic), entertainment (e.g., which movie would win the Academy Award for the
Best Picture), politics (e.g., which candidate would be elected the U.S. president),
etc. We collect all those contracts that are on binary events, regardless of their
categories, and record how each binary event had turned out.
The aim of the empirical analysis is to estimate the ex post sample proportion,
P , of event As occurrence as a function of the ex ante price, , of contract A. We
process the data in the following way. The observations are sorted by price and
evenly partitioned into a number of percentile bins, with each bin representing an
equal portion of all observations. For each percentile bin, we calculate the sample
proportion of event As occurrences whose corresponding prices fall into that bin.
Finally, we plot the sample proportions against the mid-points of the corresponding
price bins.
If the theory developed in the previous section holds, the following is what one
would expect in the empirics. Recall that the value of b depends on the distribution
of mean beliefs among traders. Since each observation in the dataset is from a certain
market with a certain distribution of mean beliefs held by the participating traders,
we can interpret each observation as a single draw from the data-generating process
5
See Appendix B for the details of the empirics.

12
associated with a certain version of Figure 1. For a price bin closer to 0, therefore,
it is more likely that the observations contained in the bin have been drawn from
the left part of Figure 1, i.e., below the crossover point. Similarly, for a price bin
closer to 1, the observations are more likely to have been drawn from the right part
of Figure 1, i.e., above the crossover point. More important, when the price bin is
near 0:5, the observations are more likely to be from around the crossover point,
suggesting a steep slope. Therefore, one would expect the empirical relationship in
a P - diagram to be close to the 45-degree line, except for (i) signicant, negative
outliers to the left of a crossover point and (ii) signicant, positive outliers to the
right.
We start with the empirical evidence from political events, one of the largest
categories in the Intrade dataset. These events, like Brexit and U.S. presidential
elections, often see a high volume of transactions between bettors. Figure 2, based
on a partition into 30 bins (i.e., 3:3% of observations per bin), plots the sample
proportion for all bins against the corresponding price. Since prices evolve over
time in the prediction markets until the random events are realized, the two panels
of the gure together capture the eect of timing by showing the estimation for two
dierent dates: (a) the rst day market opens to bettors, and (b) the last trading
day before the event is realized, respectively. In each panel, the dashed line is the
regression line, and the red dots are outliers (over 1.5 standard deviations from the
mean). Details of the regression are shown in the column 30 bins of Table 1.
A few remarks follow. First, in both panels, the regression lines fall very close to
the 45-degree line, suggesting the e ciency of markets in pricing the probabilities
of random events. The evidence of market e ciency on the rst trading day is
remarkable because, for politics, a lot of markets opened a long time sometimes
years ahead of the resolution of the events. Yet, as the regression table shows, the
slopes are statistically signicant and very close to 1, while the intercepts are not
signicant and very close to 0.
Second, although panel (a) is relatively more noisy, panel (b) shows two clear
patterns as predicted by our theory: (1) a sudden jump near price level 0:5, and (2)
signicant outliers near the crossover point. For observations in the intermediate
price range, one might think that the price is close to 0:5 because traders have solid
information suggesting an even match between outcome A and Ac . It is also
possible, however, that the market is shrouded in ambiguity as some traders, albeit
partially informed, are reluctant to trade. Just like in the example of a Knightian

13
 Figure 2: Prediction Market Data in the P - Diagram: Politics (30 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

14
 Table 1: Regression of ex post Sample Proportion on ex ante Price: Politics

ex post sample proportion

(30 bins) (50 bins)
rst trading day
ex ante price 0:950 0:953
(21:5) (22:7)
constant 0:00559 0:00645
( 0:237) ( 0:288)

last trading day

ex ante price 1:04 1:03
(39:7) (44:8)
constant 0:0228 0:0232
( 1:47) ( 1:69)
bins 30 50
t statistics in parentheses
p < 0:05, p < 0:01, p < 0:001

urn, an intermediate price in this case could mean a wide range of true probabilities.
In panel (b), for a price range between 0:45 and 0:75, the sample proportion could
be as low as 30%, or as high as 97%. A signicant jump near 0:5, therefore, is
an indication of the presence of Knightian uncertainty. As the regression table
shows, the specic patterns of outliers also cause the regression line to have a slope
larger albeit only slightly than 1.
Another important dierence between the two panels is that, in panel (b), more
observations are clustered near price levels 0 and 1. This means, by the last day,
more traders hold (posterior) beliefs that some outcome either A or Ac is very
likely to be realized, suggesting a decrease in risks over time. Such a decrease in
risks can be a result of information acquisition by the traders, who, until the random
events resolve, may have the incentives to learn about the events and update their
bets accordingly. Since risks have decreased while ambiguity remains, our empirical
evidence provides an observational distinction between the concepts of risk and
Knightian uncertainty.
Furthermore, the above empirical patterns are robust against the choice of the
number of bins. Figure 3 reproduces the diagram by partitioning the data into 50
bins instead. The column 50 bins of Table 1 shows the details of the regression.

15
Overall, the observations drawn earlier still hold.6
The observation that ambiguity remains until the last trading day suggests,
unlike risks, Knightian uncertainty may be not learnablein practice to the traders.
The intuition can be illustrated using the Knightian urn where the composition of
black and red balls is unknown. Imagine two dierent scenarios. In the rst scenario,
a subject observed repeated draws from the same Knightian urn. In this case, the
sample proportion over time would reveal the true composition of the two colors
because, after all, the composition is xed over time. In the second scenario, there
was an experimenter who replaced the Knightian urn with a new one every time
a ball was drawn by the subject. In this case, the sample proportion may not
inform the subject of what to expect in the next Knightian urn, simply because
the composition of black and red balls in the new urn could be anything of the
experimenters choosing. If the underlying data-generating process that is, the
way the experimenter changed every other Knightian urn was not learnable to the
subject, then the degree of ambiguity would not decrease over time.7 One might
think that, in politics, it is intuitively easy for traders to acquire knowledge from
polls, news reports, political analyses, etc. Yet, our empirical evidence, which is
based on a large number of prediction markets about various political events, seems
to t the second scenario, suggesting that the Knightian uncertainty of politics may
indeed be not learnable through information acquisition.
We now turn to another major category: entertainment events, such as the
winners of cinematic awards or the box o ces of movies. Figures 4 and 5 reproduce
the P - diagram for 30 bins and 50 bins, respectively, and Table 2 reports the details
of the regressions. Although qualitatively similar, the patterns are less pronounced
compared to politics. The jump near 0:5 is less clear and, interestingly, the clustering
near 0 and 1 is less marked. This evidence suggests less learning in entertainment
than in politics, which is understandable since it is more di cult for bettors to
acquire information about the general publics personal tastes of movies and music.
Politics and entertainment together accounts for over 80% of the Intrade dataset.
However, for completeness, we reproduce the empirical evidence with the full sample,
as shown in Figures 6 and 7, as well as Table 3. The patterns, essentially by
construction, are similar to what we establish above.
6
We have checked other variations between 30 and 50, which yield similar results (omitted to
limit space). Obviously, the number of bins should be neither too small (which would leave too few
points in the diagram), nor too large (which would leave too few observations per bin).
7
See Epstein and Schneider (2007) for a theoretical treatment of learning under ambiguity.

16
 Figure 3: Prediction Market Data in the P - Diagram: Politics (50 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

17
Figure 4: Prediction Market Data in the P - Diagram: Entertainment (30 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

18
Figure 5: Prediction Market Data in the P - Diagram: Entertainment (50 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

19
Table 2: Regression of ex post Sample Proportion on ex ante Price: Entertainment

ex post sample proportion

(30 bins) (50 bins)
rst trading day
ex ante price 0:936 0:933
(23:3) (20:9)
constant 0:0107 0:0115
( 0:542) ( 0:522)

last trading day

ex ante price 1:04 1:04
(37:8) (33:7)
constant 0:0438 0:0438
( 2:95) ( 2:63)
bins 30 50
t statistics in parentheses
p < 0:05, p < 0:01, p < 0:001

Table 3: Regression of ex post Sample Proportion on ex ante Price: Full Sample

ex post sample proportion

(30 bins) (50 bins)
rst trading day
ex ante price 0:929 0:917
(27:7) (31:6)
constant 0:00409 0:00144
( 0:243) ( 0:0980)

last trading day

ex ante price 1:03 1:03
(46:5) (56:6)
constant 0:0237 0:0233
( 1:94) ( 2:33)
bins 30 50
t statistics in parentheses
p < 0:05, p < 0:01, p < 0:001

20
 Figure 6: Prediction Market Data in the P - Diagram: Full Sample (30 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

21
 Figure 7: Prediction Market Data in the P - Diagram: Full Sample (50 bins).

(Note: The dashed lines are regression lines. The red dots are outliers, i.e., observations
whose residuals are more than 1.5 standard deviations from the mean.)

22
5 Concluding Remarks
Knightian uncertainty an important concept in the theoretical literature has never
been directly observed. In this paper, we have developed a model of a prediction
market with ambiguity, where traders have maxmin preferences. We have derived
more direct, observational implications of the presence of Knightian uncertainty.
Using the historical data from Intrade, we have further presented some high-level
evidence that is consistent with the prediction of our model. In particular, for price
levels close to 0:5, the market-implied, ex ante probability of a random event is not
indicative of the ex post sample proportion, suggesting the presence of Knightian
uncertainty.
Moreover, our empirical evidence has shown that, although traders seem to have
acquired information which leads to a decrease in risks, ambiguity remains until the
last trading day, suggesting that the Knightian uncertainty of the world may be not
learnableto traders. By comparing political events and entertainment events, we
have also shown that the empirical patterns we identied are more pronounced in
politics than in entertainment.
The evidence we have provided is only preliminary, since the empirics of this
paper are based on a single prediction platform that is skewed towards political and
entertainment events. In a future, empirical study, we will collect more prediction-
market data across dierent platforms and dierent event types, and we will examine
more closely the relationship between the ex post sample proportion and the ex ante
price by taking into account the type of events, the time ahead of the resolution of
randomness, and other aspects of the betting markets.

23
References
[1] Asch, Peter, Burton G. Malkiel, and Richard E. Quandt (1982). Racetrack
Betting and Informed Behavior,Journal of Financial Economics, 10(2): 187
194.

[2] Berg, Joyce, Robert Forsythe, Forrest Nelson, and Thomas Rietz (2008). Re-
sults from a Dozen Years of Election Futures Markets Research," in Handbook
of Experimental Economics Results. Charles R. Plott and Vernon L. Smith, eds.
Amsterdam; Oxford: North Holland.

[3] Caballero, Ricardo J. and Arvind Krishnamurthy (2008). Collective Risk Man-
agement in a Flight to Quality Episode,Journal of Finance, 63(5): 21952230.

[4] Caballero, Ricardo J. and Alp Simsek (2013). Fire Sales in a Model of Com-
plexity, Journal of Finance, 68(6): 25492587.

[5] Cain, Michael, David Law, and David Peel (2000). The Favourite-Longshot
Bias and Market E ciency in UK Football Betting,Scottish Journal of Polit-
ical Economy, 47(1): 2536.

[6] Dow, James and Srgio Ribeiro da Costa Werlang (1992). Uncertainty Aver-
sion, Risk Aversion, and the Optimal Choice of Portfolio,Econometrica, 60(1):
197204.

[7] Ellsberg, Daniel (1961) Risk, Ambiguity, and the Savage Axioms, Quarterly
Journal of Economics, 75(4): 643669.

[8] Epstein, Larry G. and Martin Schneider (2007). Learning under Ambiguity,
Review of Economic Studies, 74(4): 12751303.

[9] Epstein, Larry G. and Martin Schneider (2010). Ambiguity and Asset Mar-
kets, Annual Review of Financial Economics, 2: 315346.

[10] Gilboa, Itzhak and David Schmeidler (1989). Maxmin Expected Utility with
Non-Unique Prior, Journal of Mathematical Economics, 18(2): 141153.

[11] Guidolin, Massimo and Francesca Rinaldi (2010). A Simple Model of Trading
and Pricing Risky Assets under Ambiguity: Any Lessons for Policy-Makers?
Applied Financial Economics, 20(1-2): 105135.

24
[12] Huber, Peter J. and Elvezio M. Ronchetti (2009). Robust Statistics. 2nd ed.
Hoboken, N.J.: Wiley.

[13] Ipeirotis, Panos (2013). Intrade-Archive. GitHub repository,

https://github.com/ipeirotis/Intrade-Archive.

[14] Kahneman, Daniel and Amos Tversky (1979). Prospect Theory: An Analysis
of Decision under Risk, Econometrica, 47(2): 263292.

[15] Knight, Frank H. (1921). Risk, Uncertainty and Prot. Boston; New York:
Houghton Mi- in Company.

[16] Manski, Charles F. (2006). Interpreting the Predictions of Prediction Mar-

kets, Economics Letters, 91(3): 425429.

[17] Routledge, Bryan R. and Stanley E. Zin (2009). Model Uncertainty and Liq-
uidity, Review of Economic Dynamics, 12(4): 543566.

[18] Shin, Hyun Song (1992). Prices of State Contingent Claims with Insider
Traders, and the Favourite-Longshot Bias,Economic Journal, 102(411): 426
435.

[19] Snowberg, Erik and Justin Wolfers (2010). Explaining the Favorite-Long Shot
Bias: Is it Risk-Love or Misperceptions?Journal of Political Economy, 118(4):
723746.

[20] Sutherland, Wilson A. (1975). Introduction to Metric and Topological Spaces.

Oxford: Clarendon Press.

[21] Thaler, Richard H. and William T. Ziemba (1988). Anomalies: Parimutuel

Betting Markets: Racetracks and Lotteries,Journal of Economic Perspectives,
2(2): 161174.

[22] Uhlig, Maxim (2013). Ination Ambiguity and the Term Structure of U.S.
Government Bonds, Journal of Monetary Economics, 60(2): 295309.

[23] Wolfers, Justin and Eric Zitzewitz (2006). Interpreting Prediction Market
Prices as Probabilities, NBER Working Paper, No. 12200.

[24] Weitzman, Martin (1965). Utility Analysis and Group Behavior: An Empirical
Study, Journal of Political Economy, 73(1): 1826.

25
A Proofs
Proof of Proposition 1. Note that any equilibrium price has to satisfy (i)
> 2 and (ii) < 1 2 . If (i) does not hold, then qi + for all i, which means
any trader will have either a long position or a zero position not an equilibrium.
Similarly, if (ii) does not hold, no trader will have a long position, which cannot be
an equilibrium either.
Substitute (2) into (3) and rewrite the aggregate demand as
Z Z 1
q+ q
XF ( ) = wdF (q) + wdF (q) :
(1 ) + (1 )

Hence, XF ( ) = 0 if and only if

Z Z 1
(q + ) dF (q) + (q ) dF (q) = 0
+
Z Z 1 Z Z 1
, (q ) dF (q) + (q ) dF (q) + dF (q) dF (q) = 0
+ +
Z +
, EF (q) (q ) dF (q) + [F ( ) + F( + ) 1] = 0
Z +
+
, EF (q) + F (q) dq [(q ) F (q)] + [F ( ) + F( + ) 1] = 0;

where the last step follows from integration by parts. Simplifying and rearranging
terms yields the stated expression in the proposition.

Proof of Proposition 2. Let G be the space of distribution functions over

[ ;1 ], endowed with the Lvy metric `, where

` (G1 ; G2 )
inf f" > 0 j G1 (q ") " G2 G1 (q + ") + " for all q 2 [ ; 1 ]g

for any G1 ; G2 2 G. Let F be the subset of G that satises F = EF (q) for any
F 2 F. Since the Lvy metric metrizes the weak topology,8 the proposition is
equivalent to the claim that F is nowhere dense in (G; `).
Note that F is closed. Since a set is nowhere dense if and only if the complement
8
See, e.g., Huber and Ronchetti (2009), p. 28.

26
of its closure is dense,9 , it remains to be shown G n F is dense, that is, for any point
in G, there is a sequence from G n F converging to that point. It is thus enough
to show, for any F 2 F and any > 0, there exists some G 2 G n F such that
` (F; G) < .
F is non-decreasing since it is a distribution function. It follows that

lim F (q) F (EF (q) ):

q![EF (q)+ ]

We show prove the results by examining two cases.

Case 1: limq![EF (q)+ ] F (q) > F (EF (q) ).

Given > 0, we construct a distribution function G from F as

8
>
< F (q) if q 2 [ ; EF (q) 1) ;
G (q) F (EF (q) ) if q 2 [EF (q) 1 ; EF (q) + + 2) ;
>
:
F (q) if q 2 [EF (q) + + 2; 1 ];

where 1; 2 > 0 are such that function g G F satises conditions

Z EF (q)+ + 2
g (q) dq = 0
EF (q) 1

and
max fg (EF (q) 1) ; g (EF (q) + + 2 )g = :
2
It is easily veried that G is a mean-preserving spread of F , with two new atoms
created at points EF (q) 1 and EF (q) + + 2. By construction, this implies
that
Z EG (q)+ Z EF (q)+
G (q) dq = G (q) dq
EG (q) EF (q)
Z EF (q)+
= [F (q) + g (q)] dq
EF (q)
Z EF (q)+
= + g (q) dq < ;
EF (q)

where the last equality holds because F 2 F, and the inequality is due to g (EF (q) + ) <
9
See, e.g., Sutherland (1975), p. 64.

27
 R EF (q)+ R EG (q)+
0 which implies EF (q) g (q) dq < 0. Since EG (q) G (q) dq < , G 2 G nF. Finally,
let be the uniform metric, that is,

(G1 ; G2 ) sup fjG1 (q) G2 (q)j j q 2 [ ; 1 ]g

for any G1 ; G2 2 G. By construction, (F; G) = 2. Since the Lvy metric is

bounded by the uniform metric from above, that is, ` (G1 ; G2 ) (G1 ; G2 ) for any
G1 ; G2 2 G, we have ` (F; G) 2 < .

Case 2: limq![EF (q)+ ] F (q) = F (EF (q) ).

Given > 0, we construct a distribution function H from F as

8
>
< F (q) if q 2 [ ; EF (q) );
H (q) F (EF (q) )+ 3 if q 2 [EF (q) ; EF (q) + + 4) ;
>
:
F (q) if q 2 [EF (q) + + 4; 1 ];

where 3; 4 > 0 are such that function h H F satises conditions

Z EF (q)+ + 4
h (q) dq = 0
EF (q)

and
max f 3 ; h (EF (q) + + 4 )g = :
2
It is easily veried that H is a mean-preserving spread of F , with two new atoms
created at points EF (q) and EF (q) + + 4. By construction, this implies that
Z EG (q)+ Z EF (q)+
H (q) dq = H (q) dq
EG (q) EF (q)
Z EF (q)+
= [F (q) + h (q)] dq
EF (q)
Z EF (q)+
= + h (q) dq
EF (q)
= +2 3 > ;

where the last but second equality holds because F 2 F, and the last equality follows
R E (q)+
from the construction of H. Since EGG(q) H (q) dq > , H 2 G n F. Finally, similar
to Case 1, we have (F; H) = 2 and, hence, ` (F; H) < .

28
Proof of Proposition 3. Decompose XF ( ) into the aggregate supply (shorts)
SF ( ) and the aggregate demand (longs) DF ( ), where
Z Z 1
q+ q
SF ( ) = wdF (q) ; DF ( ) = wdF (q) ;
(1 ) + (1 )

and SF ( F) = DF ( F) in equilibrium. We show that an increase in shifts the

supply curve inwards. That is,
Z
dSF ( ) + @ q+
=0+ wdF ( ) wdF (q) < 0:
d (1 ) @ (1 )

dDF ( )
Similarly, an increase in shifts the demand curve inwards (i.e., d < 0). It
follows that the equilibrium quantity of trade SF ( F ), or DF ( F ) has to be
smaller as the degree of ambiguity increases.
Rq
Proof of Proposition 4. Let denote the integral of F , i.e., (q) F (q 0 ) dq 0 .
It follows from the denition of F that
Z q (
0 0 (1 m) (q) if q 2 [ ; p) ;
(q) = F q dq =
(1 m) (q) + m (q p) if q 2 [p; 1 ];

where is the integral of F . The equilibrium condition becomes

= EF (q) + ( + ) ( )
= 1 2 (1 )+ ( + ) ( );

where the second equality follows from integration by parts. Since (q) has a kink
at point p, the equilibrium price depends on the position of p relative to + and
.

Case 1: p + .

The equilibrium condition is rewritten as

= 1 2 (1 m) (1 ) m (1 p)
+ (1 m) ( + ) + m( + p) (1 m) ( ):

29
Rearranging terms and dividing both sides by 1 m yields

( + )+ ( )=1 2 (1 ):

Case 2: p > + .

The equilibrium condition is rewritten as

= 1 2 (1 m) (1 ) m (1 p)
+ (1 m) ( + ) (1 m) ( ):

Rearranging terms yields

(p )m
( + )+ ( )=1 2 (1 )+ : (4)
1 m 1 m

Note that the left-hand side of equation (4) is strictly increasing in . Thus, the
solution to the equation is a continuous and strictly increasing function of p.
Furthermore, as p ! b + , where b is the equilibrium price in Case 1, the right-
hand side of equation (4) converges to 1 2 (1 ) + 1bmm , and the solution to
the equation converges to b. In other words, the equilibrium price is continuous at
point p = b + .
Next, we show (1 )<1 2 , which implies (1 )<1 in part 1 of
the proposition. Let LHS ( ) and RHS (p) denote the left- and right-hand sides of
equation (4), as functions of and p, respectively. Note that

1 2
LHS (1 2 ) RHS (1 ) = (1 )+ (1 3 )
1 m
(1 2 )m
1 2 (1 )+
1 m
= (1 3 ) > 0:

Since LHS is strictly increasing in , the solution to the equation when p = 1

must be smaller than 1 2 .

Case 3: p < .

30
 The equilibrium condition is rewritten as

= 1 2 (1 m) (1 ) m (1 p)
+ (1 m) ( + ) + m( + p)
(1 m) ( ) m( p) :

Rearranging terms yields

(p + ) m
( + )+ ( )=1 2 (1 )+ : (5)
1 m 1 m

Similar to Case 2, the solution to equation (5) is continuous and strictly increasing
in p, and it converges to b as p ! b . Hence, the equilibrium price is continuous
at point p = b as well.
Next, we show ( ) > 2 , which implies ( )> in part 1 of the proposition.
Again, let LHS ( ) and RHS (p) denote the left- and right-hand sides of equation
(5). Note that

2
LHS (2 ) RHS ( ) = (3 ) + ( )
1 m
2 m
1 2 (1 )+
1 m
= (1 ) (3 ) [(1 ) 3 ] < 0;

where the last inequality holds because is the integral of distribution function F
over [ ; 1 ]. Since LHS is strictly increasing in , the solution to the equation
when p = must be larger than 2 .

Proof of Corollary 5. Recall that b is identied by equation

b (b + ) + (b )=1 2 (1 ):

The symmetry of F implies (1 x) = (x ) (x) for any x 2 [ ; 1 ]. Thus,

the equilibrium condition becomes

b b+ (1 b ) + (b )=1 2 1 2 + ( )
, (1 b ) (b )= ( ) = 0;

31
to which b = 0:5 is the only solution.

32
B Empirics
The historical data of Intrade was archived by Ipeirotis (2013) and is available
on GitHub. Table 4 lists all the categories of events and the number of markets
within each category. We complete the dataset by creating an outcome variable and
recording how each random event had turned out. The outcome equals 1 if an event
occurs, and it equals 0 if its complement event occurs.
Some markets have correlated outcomes, because they are about the same, un-
certain circumstances. For example, concerning the 2012 U.S. Republican Party
presidential nominee, there are 53 separate markets corresponding to 53 possible
winners, including Mitt Romney, Rick Santorum, Ron Paul, Newt Gingrich, and
any other individual not specied by the prediction platform. To avoid such cor-
relation in the observations, for each group of these correlated markets, we randomly
select one market into the aggregate sample and disregard the rest.
The total number of selected markets included in the nal analysis also shown in
Table 5. The table lists the number of observations the total as well as the number
of observations per percentile bin for political events, entertainment events, and
the full sample. The dataset is skewed towards political and entertainment events,
as the two categories together accounts for 82% of the full sample.

33
Table 4: Intrade Data: Event Categories and Number of Markets.

Event category Number of markets

Art 60
Business 43
Chess 52
Climate & Weather 861
Construction & Engineering 9
Current Events 1540
Education 1
Entertainment 8715
Fine Wine 5
Foreign Aairs 87
Legal 310
Media 10
Politics 5460
Real Estate 2
Science 20
Social & Civil 30
Technologies 65
Transportation 11

Table 5: Intrade Data: Number of Observations in Final Analysis.

Event category Total observations Observations per bin

(30 bins) (50 bins)
Politics 897 30 18
Entertainment 1157 39 23
Full sample 2509 84 50

34